RANDOMIZED COMPLETE BLOCK DESIGN (CBD)

advertisement
STAT 512
Complete Block Designs
1
RANDOMIZED COMPLETE BLOCK DESIGN (CBD)
• As with CRD, t unstructured treatments
• b blocks, each of size t, one unit per treatment per block
• Pairs of units from the same block are thought to be “more alike”
or more homogeneous, than pairs from different blocks
• N = bt total units and observations
treat. 1
treat. 2
...
treat. t
block 1
y1,1
y1,2
...
y1,t
block 2
y2,1
y2,2
...
y2,t
...
...
...
...
...
block b
yb,1
yb,2
...
yb,t
• Ordinarily reflects model and assumptions for 2-way ANOVA
STAT 512
Complete Block Designs
• Sometimes blocking is necessary, e.g. biology studies in which
animals from the same litter are more alike than animals from
different litters, but more than one litter is needed.
• Sometimes blocking is done by choice, e.g. industrial experiments
where lab-scale properties of materials give more homogeneous
results than plant-scale, but more runs are needed than can be
produced in one lab-scale “batch”.
• Either way, the point is to COMPARE TREATMENTS across
homogeneous units (to maximize experimental control).
• What’s lacking here? True REPLICATION ... i.e. every two data
values differ by at least one of block or treatment
• Some blocked designs do contain replication ... more than one
unit receiving the same treatment in the same block ... but we’re
not talking about that here.
2
STAT 512
Complete Block Designs
3
• Suppose each block-and-treatment combine in a unique way to
produce a general (unrestricted) mean:
yi,j = µi,j + i,j , i = 1...b, j = 1...t
• No replication, one mean-parameter per observation, no ability to
estimate σ:
source
Model|Mean
df
bt − 1
Residual
0
Mean Correction
1
Uncorrected Total
bt
STAT 512
Complete Block Designs
• No formal inference is possible unless some additional assumptions
are made ... d.f. “moved” from model to residual
• The usual model assumes additive structure between blocks and
treatment effects (i.e. no block-by-treatment interaction). Can
check (to some degree) with, e.g.:
– Tukey’s “one degree of freedom test”
– Simple graphics using yi,j − yi,j 0 − yi0 ,j + yi0 ,j 0
∗ “4 corners” contrasts from 2-way table
∗ Should have E = 0, V ar = 4σ 2 ...
b
t
∗ 2 × 2 of these ... a very large number, and clearly not
independent
∗ Usually looking for atypical values with common i or j ...
4
STAT 512
Complete Block Designs
Notes:
• The above remark about the “usual model assumes ... no interaction”
is based on the blocks being treated as fixed effects.
• Alternatively, if blocks are regarded as random effects, and the
block×treatment interaction (also a random effect) is included, the
appropriate denominator SS is associated with this interaction.
• Note that the arithmetic is actually the same either way ... the
(b − 1)(t − 1) degree-of-freedom denominator can be interpretted as
representing:
– residual, in a model with fixed blocks and no interaction
– (random) interaction, in a model with random blocks
• In more complex designs, the decision to call blocks “fixed” or
“random” is actually more important than you may think. We will
come back to this; for now, blocks are “fixed” unless otherwise stated.
5
STAT 512
Complete Block Designs
Now consider models:
• yi,j = βi + τj + i,j or equivalently α + βi + τj + i,j
• i = 1...b,
j = 1...t
• Ordering observations by block, then treatment within block


...
I
1 1


 1
1 ...
I 


X=

 ... ... ... ... ... ... 


1
... 1 I
• This is the model matrix for the second model, drop first column
for the first:
6
STAT 512
Complete Block Designs

1



X=
 ...


...
I
1
...
...
...

I 


... ... 

1 I
...
• There is one linear dependency among columns with first model,
or two with second
• Inferences of interest center on τ , regarding (α, β) as nuisance
parameters ... Set up a partitioned model:
– put τ -columns in X2
– put β-columns (and optionally the α-column) in X1
7
STAT 512
Complete Block Designs
LEAST-SQUARES ESTIMATION OF τ in PARTITIONED MODEL
• y = Xθ + = X1 β + X2 τ + • Here, β = (α, β1 , β2 , ...βb ), or without α
• Full normal equations:
(X0 X)θ̂ = X0 y
• Reduced normal equations:
X02 (I − H1 )X2 τ̂ = X02 (I − H1 )y
8
STAT 512
Complete Block Designs
9
• H1 is similar to HA for a CRD, both describing one-way
classification:


Jt×t
0
...
0


 0

J
...
0
t×t


H1 = 1t 
(b diagonal blocks)

 ...
...
...
... 


0
0
... Jt×t
since the blocking structure divides data into b groups of size t.
• So,
X02 H1 =
1
t
Jt×t
Jt×t
... Jt×t
because X02 = (I, I, I, ..., I), b times.
= 1t Jt×bt
STAT 512
Complete Block Designs
• What if α had been the ONLY nuisance parameter? (i.e. CRD)
X1 = 1N
H1 =
1
N JN ×N
• Then,
X02 H1 =
1
N (bJt×N )
= 1t Jt×bt
• The SAME THING ... RNE’s are the same for CBD’s and CRD’s,
0 τ is computed as if there are no blocks
d
so c
• This is a result of orthogonality; each treatment-block
combination appears the same number of times (once)
10
STAT 512
Complete Block Designs
More about this:
• The RNE’s are the same for a CBD and CRD (where ni = b)
because, apart from a rearrangement of the rows in the model
matrix:
– X2 is the same for each design (e.g. any given treatment
receives the same number of units in each design), and
– H1 X2 is the same for each design
• The book says two designs satisfy “Condition E” for “equivalent”
when they have this relationship.
• Other blocking arrangements are also “equivalent” to CRD’s in
this sense, e.g. “augmented” block designs ... see discussion in
Chapter 4.
11
STAT 512
Complete Block Designs
• So, for either CRD or CBD
X02 (I − H1 )X2 τ̂ = X02 (I − H1 )y
(X02 X2 − 1t Jt×N X2 )τ̂ = X02 y − 1t Jt×N y




y.1
y..




 y 
 y 
 .2  1  .. 
(bI − bt Jt×t )τ̂ = 
− 

 ...  t  ... 




y.t
y..

 

ȳ.1
ȳ..

 

 ȳ   ȳ 
 .2   .. 
τ̂ − τ̂¯1 = 
−

 ...   ... 

 

ȳ.t
ȳ..
τˆj − τ̂¯ = ȳ.j − ȳ..
12
STAT 512
Complete Block Designs
• Terms on left are non-unique because the same constant can be
added or subtracted from each τˆj in one solution to get another
• Can restrict the system (select a specific solution) by requiring:
– τ̂¯ = 0 → τˆj = ȳ.j − ȳ..
– τˆt = 0
→
τˆj = ȳ.j − ȳ.t
... this is really equivalent to selecting a specific g-inverse
P
P
• Estimable functions are
cj τj such that
cj = 0
P
P 21
2
• For any estimable function, V ar[ cj τˆj ] = σ
cj b ... same
formula for CRD or CBD
13
STAT 512
Complete Block Designs
So, if the form of the least-squares estimates is the same for both
designs, why is the CBD better (at least sometimes)?
• Suppose represents primarily differences in physical units ...
• CRD:
– Selecting a larger sample of units requires drawing from a less
homogeneous population (with a larger variance)
2
– N = bt units drawn from a population with V ar[] = σCRD
• CBD:
– Selecting smaller samples of units allows drawing from more
homogeneous populations (with a smaller variance)
– b groups of t units, each drawn from a (different) population
2
with V ar[] = σCBD
14
STAT 512
Complete Block Designs
• e.g. can perhaps choose between selecting:
– 50 litters of animals, 6 animals from each litter
– 300 animals from the general population
• or:
– mix 5 batches of chemicals in lab on each of 20 days
– pick 100 batches of chemicals from production facility
15
STAT 512
Complete Block Designs
P
Expected squared length of CI for
cj τj :
q
P 2
• CI is: estimate ± t × M SE × cj /b
q
P 2
• Interval length is 2 × t × M SE × cj /b
√
• Expected squared length (to avoid E M SE) is:
P 2
α
2
2
– CRD: 4t (1 − 2 , N − t)σCRD ( cj /b)
P 2
α
2
2
– CBD: 4t (1 − 2 , N − b − t + 1)σCBD ( cj /b)
• CBD quantity is smaller if:
σCBD /σCRD < t(1 − α2 , N − t)/t(1 − α2 , N − b − t + 1)
and in this case, CI’s are “tighter” on average than for CRD.
• t-ratio is only slightly < 1 unless t and b are small
16
STAT 512
Complete Block Designs
Power of F-test for τ1 = τ2 = ... = τt . Briefly,
2
• E[ SSE( HypA ) ] = σCBD
(N − b − t + 1)
• E[ SSE( Hyp0 ) ] = (β 0 X01 + τ 0 X02 )(I − H1 )(X1 β + X2 τ )
2
+ σCBD
(N − b)
– d.f. for both SSE’s are different than with CRD
– H1 is different here than with CRD, but
– X1 terms drop out for the same reason as before, and
– X02 H1 is as with the CRD, so
• Quadratic form is still
Q2|1 (τ ) = τ
as with CRD
0
X02 (I
− H1 )X2 τ =
Pt
2
b(τ
−
τ
¯
)
j
.
j=1
17
STAT 512
Complete Block Designs
• [SST/(t − 1)]/[SSE(HypA )/(N − b − t + 1)] ∼
2
F 0 (t − 1, N − b − t + 1, Q2|1 (τ )/σCBD
)
2
• Compare to F 0 (t − 1, N − t, Q2|1 (τ )/σCRD
) for CRD ... for
candidate values of σ’s and τ ’s
• In either case, the power is the probability of seeing a larger F
than the critical value determined under the central F distribution
... calculations with SAS or R as discussed before for CRD
• Note that the critical value will be somewhat smaller for the CRD
due to an increase in denominator d.f. ... Other things being
equal, this would give the CRD slightly more power, but this is
usually offset if σCBD is even slightly less than σCRD
18
Download